Tom Stellard | 75aadc2 | 2012-12-11 21:25:42 +0000 | [diff] [blame] | 1 | //===-- AMDILISelDAGToDAG.cpp - A dag to dag inst selector for AMDIL ------===// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
| 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
| 7 | // |
| 8 | //==-----------------------------------------------------------------------===// |
| 9 | // |
| 10 | /// \file |
| 11 | /// \brief Defines an instruction selector for the AMDGPU target. |
| 12 | // |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | #include "AMDGPUInstrInfo.h" |
| 15 | #include "AMDGPUISelLowering.h" // For AMDGPUISD |
| 16 | #include "AMDGPURegisterInfo.h" |
| 17 | #include "AMDILDevices.h" |
| 18 | #include "R600InstrInfo.h" |
| 19 | #include "llvm/ADT/ValueMap.h" |
| 20 | #include "llvm/CodeGen/PseudoSourceValue.h" |
| 21 | #include "llvm/CodeGen/SelectionDAGISel.h" |
| 22 | #include "llvm/Support/Compiler.h" |
Tom Stellard | 365366f | 2013-01-23 02:09:06 +0000 | [diff] [blame] | 23 | #include "llvm/CodeGen/SelectionDAG.h" |
Tom Stellard | 75aadc2 | 2012-12-11 21:25:42 +0000 | [diff] [blame] | 24 | #include <list> |
| 25 | #include <queue> |
| 26 | |
| 27 | using namespace llvm; |
| 28 | |
| 29 | //===----------------------------------------------------------------------===// |
| 30 | // Instruction Selector Implementation |
| 31 | //===----------------------------------------------------------------------===// |
| 32 | |
| 33 | namespace { |
| 34 | /// AMDGPU specific code to select AMDGPU machine instructions for |
| 35 | /// SelectionDAG operations. |
| 36 | class AMDGPUDAGToDAGISel : public SelectionDAGISel { |
| 37 | // Subtarget - Keep a pointer to the AMDGPU Subtarget around so that we can |
| 38 | // make the right decision when generating code for different targets. |
| 39 | const AMDGPUSubtarget &Subtarget; |
| 40 | public: |
| 41 | AMDGPUDAGToDAGISel(TargetMachine &TM); |
| 42 | virtual ~AMDGPUDAGToDAGISel(); |
| 43 | |
| 44 | SDNode *Select(SDNode *N); |
| 45 | virtual const char *getPassName() const; |
| 46 | |
| 47 | private: |
| 48 | inline SDValue getSmallIPtrImm(unsigned Imm); |
Tom Stellard | 365366f | 2013-01-23 02:09:06 +0000 | [diff] [blame] | 49 | bool FoldOperands(unsigned, const R600InstrInfo *, std::vector<SDValue> &); |
Tom Stellard | 75aadc2 | 2012-12-11 21:25:42 +0000 | [diff] [blame] | 50 | |
| 51 | // Complex pattern selectors |
| 52 | bool SelectADDRParam(SDValue Addr, SDValue& R1, SDValue& R2); |
| 53 | bool SelectADDR(SDValue N, SDValue &R1, SDValue &R2); |
| 54 | bool SelectADDR64(SDValue N, SDValue &R1, SDValue &R2); |
| 55 | |
| 56 | static bool checkType(const Value *ptr, unsigned int addrspace); |
| 57 | static const Value *getBasePointerValue(const Value *V); |
| 58 | |
| 59 | static bool isGlobalStore(const StoreSDNode *N); |
| 60 | static bool isPrivateStore(const StoreSDNode *N); |
| 61 | static bool isLocalStore(const StoreSDNode *N); |
| 62 | static bool isRegionStore(const StoreSDNode *N); |
| 63 | |
| 64 | static bool isCPLoad(const LoadSDNode *N); |
| 65 | static bool isConstantLoad(const LoadSDNode *N, int cbID); |
| 66 | static bool isGlobalLoad(const LoadSDNode *N); |
| 67 | static bool isParamLoad(const LoadSDNode *N); |
| 68 | static bool isPrivateLoad(const LoadSDNode *N); |
| 69 | static bool isLocalLoad(const LoadSDNode *N); |
| 70 | static bool isRegionLoad(const LoadSDNode *N); |
| 71 | |
Tom Stellard | 365366f | 2013-01-23 02:09:06 +0000 | [diff] [blame] | 72 | bool SelectGlobalValueConstantOffset(SDValue Addr, SDValue& IntPtr); |
| 73 | bool SelectGlobalValueVariableOffset(SDValue Addr, |
| 74 | SDValue &BaseReg, SDValue& Offset); |
Tom Stellard | 75aadc2 | 2012-12-11 21:25:42 +0000 | [diff] [blame] | 75 | bool SelectADDR8BitOffset(SDValue Addr, SDValue& Base, SDValue& Offset); |
| 76 | bool SelectADDRReg(SDValue Addr, SDValue& Base, SDValue& Offset); |
| 77 | bool SelectADDRVTX_READ(SDValue Addr, SDValue &Base, SDValue &Offset); |
| 78 | |
| 79 | // Include the pieces autogenerated from the target description. |
| 80 | #include "AMDGPUGenDAGISel.inc" |
| 81 | }; |
| 82 | } // end anonymous namespace |
| 83 | |
| 84 | /// \brief This pass converts a legalized DAG into a AMDGPU-specific |
| 85 | // DAG, ready for instruction scheduling. |
| 86 | FunctionPass *llvm::createAMDGPUISelDag(TargetMachine &TM |
| 87 | ) { |
| 88 | return new AMDGPUDAGToDAGISel(TM); |
| 89 | } |
| 90 | |
| 91 | AMDGPUDAGToDAGISel::AMDGPUDAGToDAGISel(TargetMachine &TM |
| 92 | ) |
| 93 | : SelectionDAGISel(TM), Subtarget(TM.getSubtarget<AMDGPUSubtarget>()) { |
| 94 | } |
| 95 | |
| 96 | AMDGPUDAGToDAGISel::~AMDGPUDAGToDAGISel() { |
| 97 | } |
| 98 | |
| 99 | SDValue AMDGPUDAGToDAGISel::getSmallIPtrImm(unsigned int Imm) { |
| 100 | return CurDAG->getTargetConstant(Imm, MVT::i32); |
| 101 | } |
| 102 | |
| 103 | bool AMDGPUDAGToDAGISel::SelectADDRParam( |
| 104 | SDValue Addr, SDValue& R1, SDValue& R2) { |
| 105 | |
| 106 | if (Addr.getOpcode() == ISD::FrameIndex) { |
| 107 | if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) { |
| 108 | R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i32); |
| 109 | R2 = CurDAG->getTargetConstant(0, MVT::i32); |
| 110 | } else { |
| 111 | R1 = Addr; |
| 112 | R2 = CurDAG->getTargetConstant(0, MVT::i32); |
| 113 | } |
| 114 | } else if (Addr.getOpcode() == ISD::ADD) { |
| 115 | R1 = Addr.getOperand(0); |
| 116 | R2 = Addr.getOperand(1); |
| 117 | } else { |
| 118 | R1 = Addr; |
| 119 | R2 = CurDAG->getTargetConstant(0, MVT::i32); |
| 120 | } |
| 121 | return true; |
| 122 | } |
| 123 | |
| 124 | bool AMDGPUDAGToDAGISel::SelectADDR(SDValue Addr, SDValue& R1, SDValue& R2) { |
| 125 | if (Addr.getOpcode() == ISD::TargetExternalSymbol || |
| 126 | Addr.getOpcode() == ISD::TargetGlobalAddress) { |
| 127 | return false; |
| 128 | } |
| 129 | return SelectADDRParam(Addr, R1, R2); |
| 130 | } |
| 131 | |
| 132 | |
| 133 | bool AMDGPUDAGToDAGISel::SelectADDR64(SDValue Addr, SDValue& R1, SDValue& R2) { |
| 134 | if (Addr.getOpcode() == ISD::TargetExternalSymbol || |
| 135 | Addr.getOpcode() == ISD::TargetGlobalAddress) { |
| 136 | return false; |
| 137 | } |
| 138 | |
| 139 | if (Addr.getOpcode() == ISD::FrameIndex) { |
| 140 | if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) { |
| 141 | R1 = CurDAG->getTargetFrameIndex(FIN->getIndex(), MVT::i64); |
| 142 | R2 = CurDAG->getTargetConstant(0, MVT::i64); |
| 143 | } else { |
| 144 | R1 = Addr; |
| 145 | R2 = CurDAG->getTargetConstant(0, MVT::i64); |
| 146 | } |
| 147 | } else if (Addr.getOpcode() == ISD::ADD) { |
| 148 | R1 = Addr.getOperand(0); |
| 149 | R2 = Addr.getOperand(1); |
| 150 | } else { |
| 151 | R1 = Addr; |
| 152 | R2 = CurDAG->getTargetConstant(0, MVT::i64); |
| 153 | } |
| 154 | return true; |
| 155 | } |
| 156 | |
| 157 | SDNode *AMDGPUDAGToDAGISel::Select(SDNode *N) { |
| 158 | unsigned int Opc = N->getOpcode(); |
| 159 | if (N->isMachineOpcode()) { |
| 160 | return NULL; // Already selected. |
| 161 | } |
| 162 | switch (Opc) { |
| 163 | default: break; |
| 164 | case ISD::FrameIndex: { |
| 165 | if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(N)) { |
| 166 | unsigned int FI = FIN->getIndex(); |
| 167 | EVT OpVT = N->getValueType(0); |
| 168 | unsigned int NewOpc = AMDGPU::COPY; |
| 169 | SDValue TFI = CurDAG->getTargetFrameIndex(FI, MVT::i32); |
| 170 | return CurDAG->SelectNodeTo(N, NewOpc, OpVT, TFI); |
| 171 | } |
| 172 | break; |
| 173 | } |
| 174 | case ISD::ConstantFP: |
| 175 | case ISD::Constant: { |
| 176 | const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>(); |
| 177 | // XXX: Custom immediate lowering not implemented yet. Instead we use |
| 178 | // pseudo instructions defined in SIInstructions.td |
| 179 | if (ST.device()->getGeneration() > AMDGPUDeviceInfo::HD6XXX) { |
| 180 | break; |
| 181 | } |
| 182 | const R600InstrInfo *TII = static_cast<const R600InstrInfo*>(TM.getInstrInfo()); |
| 183 | |
| 184 | uint64_t ImmValue = 0; |
| 185 | unsigned ImmReg = AMDGPU::ALU_LITERAL_X; |
| 186 | |
| 187 | if (N->getOpcode() == ISD::ConstantFP) { |
| 188 | // XXX: 64-bit Immediates not supported yet |
| 189 | assert(N->getValueType(0) != MVT::f64); |
| 190 | |
| 191 | ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N); |
| 192 | APFloat Value = C->getValueAPF(); |
| 193 | float FloatValue = Value.convertToFloat(); |
| 194 | if (FloatValue == 0.0) { |
| 195 | ImmReg = AMDGPU::ZERO; |
| 196 | } else if (FloatValue == 0.5) { |
| 197 | ImmReg = AMDGPU::HALF; |
| 198 | } else if (FloatValue == 1.0) { |
| 199 | ImmReg = AMDGPU::ONE; |
| 200 | } else { |
| 201 | ImmValue = Value.bitcastToAPInt().getZExtValue(); |
| 202 | } |
| 203 | } else { |
| 204 | // XXX: 64-bit Immediates not supported yet |
| 205 | assert(N->getValueType(0) != MVT::i64); |
| 206 | |
| 207 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(N); |
| 208 | if (C->getZExtValue() == 0) { |
| 209 | ImmReg = AMDGPU::ZERO; |
| 210 | } else if (C->getZExtValue() == 1) { |
| 211 | ImmReg = AMDGPU::ONE_INT; |
| 212 | } else { |
| 213 | ImmValue = C->getZExtValue(); |
| 214 | } |
| 215 | } |
| 216 | |
| 217 | for (SDNode::use_iterator Use = N->use_begin(), Next = llvm::next(Use); |
| 218 | Use != SDNode::use_end(); Use = Next) { |
| 219 | Next = llvm::next(Use); |
| 220 | std::vector<SDValue> Ops; |
| 221 | for (unsigned i = 0; i < Use->getNumOperands(); ++i) { |
| 222 | Ops.push_back(Use->getOperand(i)); |
| 223 | } |
| 224 | |
| 225 | if (!Use->isMachineOpcode()) { |
| 226 | if (ImmReg == AMDGPU::ALU_LITERAL_X) { |
| 227 | // We can only use literal constants (e.g. AMDGPU::ZERO, |
| 228 | // AMDGPU::ONE, etc) in machine opcodes. |
| 229 | continue; |
| 230 | } |
| 231 | } else { |
| 232 | if (!TII->isALUInstr(Use->getMachineOpcode())) { |
| 233 | continue; |
| 234 | } |
| 235 | |
| 236 | int ImmIdx = TII->getOperandIdx(Use->getMachineOpcode(), R600Operands::IMM); |
| 237 | assert(ImmIdx != -1); |
| 238 | |
| 239 | // subtract one from ImmIdx, because the DST operand is usually index |
| 240 | // 0 for MachineInstrs, but we have no DST in the Ops vector. |
| 241 | ImmIdx--; |
| 242 | |
| 243 | // Check that we aren't already using an immediate. |
| 244 | // XXX: It's possible for an instruction to have more than one |
| 245 | // immediate operand, but this is not supported yet. |
| 246 | if (ImmReg == AMDGPU::ALU_LITERAL_X) { |
| 247 | ConstantSDNode *C = dyn_cast<ConstantSDNode>(Use->getOperand(ImmIdx)); |
| 248 | assert(C); |
| 249 | |
| 250 | if (C->getZExtValue() != 0) { |
| 251 | // This instruction is already using an immediate. |
| 252 | continue; |
| 253 | } |
| 254 | |
| 255 | // Set the immediate value |
| 256 | Ops[ImmIdx] = CurDAG->getTargetConstant(ImmValue, MVT::i32); |
| 257 | } |
| 258 | } |
| 259 | // Set the immediate register |
| 260 | Ops[Use.getOperandNo()] = CurDAG->getRegister(ImmReg, MVT::i32); |
| 261 | |
| 262 | CurDAG->UpdateNodeOperands(*Use, Ops.data(), Use->getNumOperands()); |
| 263 | } |
| 264 | break; |
| 265 | } |
| 266 | } |
Tom Stellard | 365366f | 2013-01-23 02:09:06 +0000 | [diff] [blame] | 267 | SDNode *Result = SelectCode(N); |
| 268 | |
| 269 | // Fold operands of selected node |
| 270 | |
| 271 | const AMDGPUSubtarget &ST = TM.getSubtarget<AMDGPUSubtarget>(); |
| 272 | if (ST.device()->getGeneration() <= AMDGPUDeviceInfo::HD6XXX) { |
| 273 | const R600InstrInfo *TII = |
| 274 | static_cast<const R600InstrInfo*>(TM.getInstrInfo()); |
Tom Stellard | 4926921 | 2013-01-31 22:11:54 +0000 | [diff] [blame^] | 275 | if (Result && Result->isMachineOpcode() |
| 276 | && TII->isALUInstr(Result->getMachineOpcode())) { |
| 277 | // Fold FNEG/FABS/CONST_ADDRESS |
| 278 | // TODO: Isel can generate multiple MachineInst, we need to recursively |
| 279 | // parse Result |
Tom Stellard | 365366f | 2013-01-23 02:09:06 +0000 | [diff] [blame] | 280 | bool IsModified = false; |
| 281 | do { |
| 282 | std::vector<SDValue> Ops; |
| 283 | for(SDNode::op_iterator I = Result->op_begin(), E = Result->op_end(); |
| 284 | I != E; ++I) |
| 285 | Ops.push_back(*I); |
| 286 | IsModified = FoldOperands(Result->getMachineOpcode(), TII, Ops); |
| 287 | if (IsModified) { |
Tom Stellard | 4926921 | 2013-01-31 22:11:54 +0000 | [diff] [blame^] | 288 | Result = CurDAG->UpdateNodeOperands(Result, Ops.data(), Ops.size()); |
Tom Stellard | 365366f | 2013-01-23 02:09:06 +0000 | [diff] [blame] | 289 | } |
| 290 | } while (IsModified); |
Tom Stellard | 4926921 | 2013-01-31 22:11:54 +0000 | [diff] [blame^] | 291 | |
| 292 | // If node has a single use which is CLAMP_R600, folds it |
| 293 | if (Result->hasOneUse() && Result->isMachineOpcode()) { |
| 294 | SDNode *PotentialClamp = *Result->use_begin(); |
| 295 | if (PotentialClamp->isMachineOpcode() && |
| 296 | PotentialClamp->getMachineOpcode() == AMDGPU::CLAMP_R600) { |
| 297 | unsigned ClampIdx = |
| 298 | TII->getOperandIdx(Result->getMachineOpcode(), R600Operands::CLAMP); |
| 299 | std::vector<SDValue> Ops; |
| 300 | unsigned NumOp = Result->getNumOperands(); |
| 301 | for (unsigned i = 0; i < NumOp; ++i) { |
| 302 | Ops.push_back(Result->getOperand(i)); |
| 303 | } |
| 304 | Ops[ClampIdx - 1] = CurDAG->getTargetConstant(1, MVT::i32); |
| 305 | Result = CurDAG->SelectNodeTo(PotentialClamp, |
| 306 | Result->getMachineOpcode(), PotentialClamp->getVTList(), |
| 307 | Ops.data(), NumOp); |
| 308 | } |
| 309 | } |
Tom Stellard | 365366f | 2013-01-23 02:09:06 +0000 | [diff] [blame] | 310 | } |
| 311 | } |
| 312 | |
| 313 | return Result; |
| 314 | } |
| 315 | |
| 316 | bool AMDGPUDAGToDAGISel::FoldOperands(unsigned Opcode, |
| 317 | const R600InstrInfo *TII, std::vector<SDValue> &Ops) { |
| 318 | int OperandIdx[] = { |
| 319 | TII->getOperandIdx(Opcode, R600Operands::SRC0), |
| 320 | TII->getOperandIdx(Opcode, R600Operands::SRC1), |
| 321 | TII->getOperandIdx(Opcode, R600Operands::SRC2) |
| 322 | }; |
| 323 | int SelIdx[] = { |
| 324 | TII->getOperandIdx(Opcode, R600Operands::SRC0_SEL), |
| 325 | TII->getOperandIdx(Opcode, R600Operands::SRC1_SEL), |
| 326 | TII->getOperandIdx(Opcode, R600Operands::SRC2_SEL) |
| 327 | }; |
Tom Stellard | 4926921 | 2013-01-31 22:11:54 +0000 | [diff] [blame^] | 328 | int NegIdx[] = { |
| 329 | TII->getOperandIdx(Opcode, R600Operands::SRC0_NEG), |
| 330 | TII->getOperandIdx(Opcode, R600Operands::SRC1_NEG), |
| 331 | TII->getOperandIdx(Opcode, R600Operands::SRC2_NEG) |
| 332 | }; |
| 333 | int AbsIdx[] = { |
| 334 | TII->getOperandIdx(Opcode, R600Operands::SRC0_ABS), |
| 335 | TII->getOperandIdx(Opcode, R600Operands::SRC1_ABS), |
| 336 | -1 |
| 337 | }; |
| 338 | |
Tom Stellard | 365366f | 2013-01-23 02:09:06 +0000 | [diff] [blame] | 339 | for (unsigned i = 0; i < 3; i++) { |
| 340 | if (OperandIdx[i] < 0) |
| 341 | return false; |
| 342 | SDValue Operand = Ops[OperandIdx[i] - 1]; |
| 343 | switch (Operand.getOpcode()) { |
| 344 | case AMDGPUISD::CONST_ADDRESS: { |
| 345 | SDValue CstOffset; |
| 346 | if (!Operand.getValueType().isVector() && |
| 347 | SelectGlobalValueConstantOffset(Operand.getOperand(0), CstOffset)) { |
| 348 | Ops[OperandIdx[i] - 1] = CurDAG->getRegister(AMDGPU::ALU_CONST, MVT::f32); |
| 349 | Ops[SelIdx[i] - 1] = CstOffset; |
| 350 | return true; |
| 351 | } |
| 352 | } |
| 353 | break; |
Tom Stellard | 4926921 | 2013-01-31 22:11:54 +0000 | [diff] [blame^] | 354 | case ISD::FNEG: |
| 355 | if (NegIdx[i] < 0) |
| 356 | break; |
| 357 | Ops[OperandIdx[i] - 1] = Operand.getOperand(0); |
| 358 | Ops[NegIdx[i] - 1] = CurDAG->getTargetConstant(1, MVT::i32); |
| 359 | return true; |
| 360 | case ISD::FABS: |
| 361 | if (AbsIdx[i] < 0) |
| 362 | break; |
| 363 | Ops[OperandIdx[i] - 1] = Operand.getOperand(0); |
| 364 | Ops[AbsIdx[i] - 1] = CurDAG->getTargetConstant(1, MVT::i32); |
| 365 | return true; |
Tom Stellard | dd04c83 | 2013-01-31 22:11:53 +0000 | [diff] [blame] | 366 | case ISD::BITCAST: |
| 367 | Ops[OperandIdx[i] - 1] = Operand.getOperand(0); |
| 368 | return true; |
Tom Stellard | 365366f | 2013-01-23 02:09:06 +0000 | [diff] [blame] | 369 | default: |
| 370 | break; |
| 371 | } |
| 372 | } |
| 373 | return false; |
Tom Stellard | 75aadc2 | 2012-12-11 21:25:42 +0000 | [diff] [blame] | 374 | } |
| 375 | |
| 376 | bool AMDGPUDAGToDAGISel::checkType(const Value *ptr, unsigned int addrspace) { |
| 377 | if (!ptr) { |
| 378 | return false; |
| 379 | } |
| 380 | Type *ptrType = ptr->getType(); |
| 381 | return dyn_cast<PointerType>(ptrType)->getAddressSpace() == addrspace; |
| 382 | } |
| 383 | |
| 384 | const Value * AMDGPUDAGToDAGISel::getBasePointerValue(const Value *V) { |
| 385 | if (!V) { |
| 386 | return NULL; |
| 387 | } |
| 388 | const Value *ret = NULL; |
| 389 | ValueMap<const Value *, bool> ValueBitMap; |
| 390 | std::queue<const Value *, std::list<const Value *> > ValueQueue; |
| 391 | ValueQueue.push(V); |
| 392 | while (!ValueQueue.empty()) { |
| 393 | V = ValueQueue.front(); |
| 394 | if (ValueBitMap.find(V) == ValueBitMap.end()) { |
| 395 | ValueBitMap[V] = true; |
| 396 | if (dyn_cast<Argument>(V) && dyn_cast<PointerType>(V->getType())) { |
| 397 | ret = V; |
| 398 | break; |
| 399 | } else if (dyn_cast<GlobalVariable>(V)) { |
| 400 | ret = V; |
| 401 | break; |
| 402 | } else if (dyn_cast<Constant>(V)) { |
| 403 | const ConstantExpr *CE = dyn_cast<ConstantExpr>(V); |
| 404 | if (CE) { |
| 405 | ValueQueue.push(CE->getOperand(0)); |
| 406 | } |
| 407 | } else if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) { |
| 408 | ret = AI; |
| 409 | break; |
| 410 | } else if (const Instruction *I = dyn_cast<Instruction>(V)) { |
| 411 | uint32_t numOps = I->getNumOperands(); |
| 412 | for (uint32_t x = 0; x < numOps; ++x) { |
| 413 | ValueQueue.push(I->getOperand(x)); |
| 414 | } |
| 415 | } else { |
| 416 | assert(!"Found a Value that we didn't know how to handle!"); |
| 417 | } |
| 418 | } |
| 419 | ValueQueue.pop(); |
| 420 | } |
| 421 | return ret; |
| 422 | } |
| 423 | |
| 424 | bool AMDGPUDAGToDAGISel::isGlobalStore(const StoreSDNode *N) { |
| 425 | return checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS); |
| 426 | } |
| 427 | |
| 428 | bool AMDGPUDAGToDAGISel::isPrivateStore(const StoreSDNode *N) { |
| 429 | return (!checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS) |
| 430 | && !checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS) |
| 431 | && !checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS)); |
| 432 | } |
| 433 | |
| 434 | bool AMDGPUDAGToDAGISel::isLocalStore(const StoreSDNode *N) { |
| 435 | return checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS); |
| 436 | } |
| 437 | |
| 438 | bool AMDGPUDAGToDAGISel::isRegionStore(const StoreSDNode *N) { |
| 439 | return checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS); |
| 440 | } |
| 441 | |
| 442 | bool AMDGPUDAGToDAGISel::isConstantLoad(const LoadSDNode *N, int cbID) { |
| 443 | if (checkType(N->getSrcValue(), AMDGPUAS::CONSTANT_ADDRESS)) { |
| 444 | return true; |
| 445 | } |
| 446 | MachineMemOperand *MMO = N->getMemOperand(); |
| 447 | const Value *V = MMO->getValue(); |
| 448 | const Value *BV = getBasePointerValue(V); |
| 449 | if (MMO |
| 450 | && MMO->getValue() |
| 451 | && ((V && dyn_cast<GlobalValue>(V)) |
| 452 | || (BV && dyn_cast<GlobalValue>( |
| 453 | getBasePointerValue(MMO->getValue()))))) { |
| 454 | return checkType(N->getSrcValue(), AMDGPUAS::PRIVATE_ADDRESS); |
| 455 | } else { |
| 456 | return false; |
| 457 | } |
| 458 | } |
| 459 | |
| 460 | bool AMDGPUDAGToDAGISel::isGlobalLoad(const LoadSDNode *N) { |
| 461 | return checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS); |
| 462 | } |
| 463 | |
| 464 | bool AMDGPUDAGToDAGISel::isParamLoad(const LoadSDNode *N) { |
| 465 | return checkType(N->getSrcValue(), AMDGPUAS::PARAM_I_ADDRESS); |
| 466 | } |
| 467 | |
| 468 | bool AMDGPUDAGToDAGISel::isLocalLoad(const LoadSDNode *N) { |
| 469 | return checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS); |
| 470 | } |
| 471 | |
| 472 | bool AMDGPUDAGToDAGISel::isRegionLoad(const LoadSDNode *N) { |
| 473 | return checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS); |
| 474 | } |
| 475 | |
| 476 | bool AMDGPUDAGToDAGISel::isCPLoad(const LoadSDNode *N) { |
| 477 | MachineMemOperand *MMO = N->getMemOperand(); |
| 478 | if (checkType(N->getSrcValue(), AMDGPUAS::PRIVATE_ADDRESS)) { |
| 479 | if (MMO) { |
| 480 | const Value *V = MMO->getValue(); |
| 481 | const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V); |
| 482 | if (PSV && PSV == PseudoSourceValue::getConstantPool()) { |
| 483 | return true; |
| 484 | } |
| 485 | } |
| 486 | } |
| 487 | return false; |
| 488 | } |
| 489 | |
| 490 | bool AMDGPUDAGToDAGISel::isPrivateLoad(const LoadSDNode *N) { |
| 491 | if (checkType(N->getSrcValue(), AMDGPUAS::PRIVATE_ADDRESS)) { |
| 492 | // Check to make sure we are not a constant pool load or a constant load |
| 493 | // that is marked as a private load |
| 494 | if (isCPLoad(N) || isConstantLoad(N, -1)) { |
| 495 | return false; |
| 496 | } |
| 497 | } |
| 498 | if (!checkType(N->getSrcValue(), AMDGPUAS::LOCAL_ADDRESS) |
| 499 | && !checkType(N->getSrcValue(), AMDGPUAS::GLOBAL_ADDRESS) |
| 500 | && !checkType(N->getSrcValue(), AMDGPUAS::REGION_ADDRESS) |
| 501 | && !checkType(N->getSrcValue(), AMDGPUAS::CONSTANT_ADDRESS) |
| 502 | && !checkType(N->getSrcValue(), AMDGPUAS::PARAM_D_ADDRESS) |
| 503 | && !checkType(N->getSrcValue(), AMDGPUAS::PARAM_I_ADDRESS)) { |
| 504 | return true; |
| 505 | } |
| 506 | return false; |
| 507 | } |
| 508 | |
| 509 | const char *AMDGPUDAGToDAGISel::getPassName() const { |
| 510 | return "AMDGPU DAG->DAG Pattern Instruction Selection"; |
| 511 | } |
| 512 | |
| 513 | #ifdef DEBUGTMP |
| 514 | #undef INT64_C |
| 515 | #endif |
| 516 | #undef DEBUGTMP |
| 517 | |
| 518 | ///==== AMDGPU Functions ====/// |
| 519 | |
Tom Stellard | 365366f | 2013-01-23 02:09:06 +0000 | [diff] [blame] | 520 | bool AMDGPUDAGToDAGISel::SelectGlobalValueConstantOffset(SDValue Addr, |
| 521 | SDValue& IntPtr) { |
| 522 | if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Addr)) { |
| 523 | IntPtr = CurDAG->getIntPtrConstant(Cst->getZExtValue() / 4, true); |
| 524 | return true; |
| 525 | } |
| 526 | return false; |
| 527 | } |
| 528 | |
| 529 | bool AMDGPUDAGToDAGISel::SelectGlobalValueVariableOffset(SDValue Addr, |
| 530 | SDValue& BaseReg, SDValue &Offset) { |
| 531 | if (!dyn_cast<ConstantSDNode>(Addr)) { |
| 532 | BaseReg = Addr; |
| 533 | Offset = CurDAG->getIntPtrConstant(0, true); |
| 534 | return true; |
| 535 | } |
| 536 | return false; |
| 537 | } |
| 538 | |
Tom Stellard | 75aadc2 | 2012-12-11 21:25:42 +0000 | [diff] [blame] | 539 | bool AMDGPUDAGToDAGISel::SelectADDR8BitOffset(SDValue Addr, SDValue& Base, |
| 540 | SDValue& Offset) { |
| 541 | if (Addr.getOpcode() == ISD::TargetExternalSymbol || |
| 542 | Addr.getOpcode() == ISD::TargetGlobalAddress) { |
| 543 | return false; |
| 544 | } |
| 545 | |
| 546 | |
| 547 | if (Addr.getOpcode() == ISD::ADD) { |
| 548 | bool Match = false; |
| 549 | |
| 550 | // Find the base ptr and the offset |
| 551 | for (unsigned i = 0; i < Addr.getNumOperands(); i++) { |
| 552 | SDValue Arg = Addr.getOperand(i); |
| 553 | ConstantSDNode * OffsetNode = dyn_cast<ConstantSDNode>(Arg); |
| 554 | // This arg isn't a constant so it must be the base PTR. |
| 555 | if (!OffsetNode) { |
| 556 | Base = Addr.getOperand(i); |
| 557 | continue; |
| 558 | } |
| 559 | // Check if the constant argument fits in 8-bits. The offset is in bytes |
| 560 | // so we need to convert it to dwords. |
| 561 | if (isUInt<8>(OffsetNode->getZExtValue() >> 2)) { |
| 562 | Match = true; |
| 563 | Offset = CurDAG->getTargetConstant(OffsetNode->getZExtValue() >> 2, |
| 564 | MVT::i32); |
| 565 | } |
| 566 | } |
| 567 | return Match; |
| 568 | } |
| 569 | |
| 570 | // Default case, no offset |
| 571 | Base = Addr; |
| 572 | Offset = CurDAG->getTargetConstant(0, MVT::i32); |
| 573 | return true; |
| 574 | } |
| 575 | |
| 576 | bool AMDGPUDAGToDAGISel::SelectADDRVTX_READ(SDValue Addr, SDValue &Base, |
| 577 | SDValue &Offset) { |
| 578 | ConstantSDNode * IMMOffset; |
| 579 | |
| 580 | if (Addr.getOpcode() == ISD::ADD |
| 581 | && (IMMOffset = dyn_cast<ConstantSDNode>(Addr.getOperand(1))) |
| 582 | && isInt<16>(IMMOffset->getZExtValue())) { |
| 583 | |
| 584 | Base = Addr.getOperand(0); |
| 585 | Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), MVT::i32); |
| 586 | return true; |
| 587 | // If the pointer address is constant, we can move it to the offset field. |
| 588 | } else if ((IMMOffset = dyn_cast<ConstantSDNode>(Addr)) |
| 589 | && isInt<16>(IMMOffset->getZExtValue())) { |
| 590 | Base = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), |
| 591 | CurDAG->getEntryNode().getDebugLoc(), |
| 592 | AMDGPU::ZERO, MVT::i32); |
| 593 | Offset = CurDAG->getTargetConstant(IMMOffset->getZExtValue(), MVT::i32); |
| 594 | return true; |
| 595 | } |
| 596 | |
| 597 | // Default case, no offset |
| 598 | Base = Addr; |
| 599 | Offset = CurDAG->getTargetConstant(0, MVT::i32); |
| 600 | return true; |
| 601 | } |
| 602 | |
| 603 | bool AMDGPUDAGToDAGISel::SelectADDRReg(SDValue Addr, SDValue& Base, |
| 604 | SDValue& Offset) { |
| 605 | if (Addr.getOpcode() == ISD::TargetExternalSymbol || |
| 606 | Addr.getOpcode() == ISD::TargetGlobalAddress || |
| 607 | Addr.getOpcode() != ISD::ADD) { |
| 608 | return false; |
| 609 | } |
| 610 | |
| 611 | Base = Addr.getOperand(0); |
| 612 | Offset = Addr.getOperand(1); |
| 613 | |
| 614 | return true; |
| 615 | } |